Robot wiping learns force, not just motion: one demo becomes a contact prompt, 78.7% on unseen objects

A robot can now pick up both the motion and the interaction intensity of a contact-rich behavior from one demonstration, instead of imitating the movement alone. Earlier in-context robot learning specified tasks mainly through visual, language, or action-trajectory context; in wiping, light and heavy strokes follow similar paths, so visual and kinematic cues alone do not distinguish the required contact intensity. Across real wiping, rotation, and insertion tasks with three unseen object instances each, the method reaches 78.7% overall success against 65.7% for the strongest baseline BPP*, 40.7% for DP-FT, and 37.0% for the other baseline, taking best or tied-best results on eight of nine instances.

The mechanism compresses demonstrated motion, bilateral 6D fingertip wrenches, and distributed force maps into a time-ordered prompt, and execution retrieves from that prompt using online observations. Force and tactile signals have mostly served as feedback control or training supervision rather than as the behavioral specification of desired contact intensity and its temporal evolution. A shared 2D CNN encodes the two fingertips' force maps and a shared MLP encodes their wrenches, cross-attention fuses them with motion features, and a dilated temporal convolutional network encodes temporal evolution; the policy generates a 32-step action chunk by conditional flow matching, executes the first 16 actions, and runs inference on a single RTX 4090.

The method performs one-shot transfer within known task families: at test time the robot receives a single reference demonstration collected on an unseen object, while the prompt encoder, online observation encoder, and policy network stay frozen, and the demonstration is encoded once and cached throughout execution. Evaluation covers real wiping, rotation, and insertion tasks with three unseen object instances per family, using four reference demonstrations and three trials each per method and object. The authors state that future work will extend the approach to more diverse tasks, richer temporal interaction patterns, and different robotic platforms.

Insertion remains the harder family: the method improves on instances A and B but reaches 33.3% on instance C, below BPP* at 41.7%. In the reference-contact ablation, a position-only prompt fails to preserve the light-medium-heavy intensity ordering, indicating that graded intensity depends on contact information in the reference. The height-perturbation experiment shifts the test object vertically relative to the reference configuration, and without online contact the execution often fails to establish surface contact. The authors also note that extending to more tasks, richer temporal interaction patterns, and different robotic platforms is left to future work.